Stanford open-sources CS329Z agent course: build your own Claude Code for under $5 in API credits
stanfordnlp · x · 2026-11-12
Stanford's new course CS329Z "Engineering AI Agents" (Prof. Diyi Yang) open-sourced its assignments, with lectures posted daily on Stanford Online YouTube (lecture 1 already up). The core assignment: write an agent harness from scratch — wrapping the model to handle tools, parse responses, and execute commands. The instructor claims under $5 of API credits gets you your own "Claude Code".
Key details:
- Starter tests all fail; your job is making them pass. Non-model tests are free to rerun; model calls are cached locally so repeated requests cost nothing extra.
- Default model is openai/gpt-6-luna; external students need their own OpenRouter key.
- Assignment 2, by SWE-bench co-author John Yang, teaches writing a Terminal-Bench-style task with Harbor.
Caveats: it's a live course with a tentative schedule, lectures released daily; "$5" covers API credits only — heavy debugging costs more.
More from coding & agent
- AI-powered three.js editing is insanely fast now, dev reports incredible quality — ZeroStateReflex · 2026-11-12
- Anthropic study: per-user agents perform worse than one shared agent on group tasks — AndrewLampinen · 2026-11-12
- Devs explore home-IP agents: Tailscale tricks and a $99 headless Android box idea — Darpinian · 2026-11-12
- Dev tip: steering smaller models with tighter context beats picking the biggest model — daniel_mac8 · 2026-11-12
- Standard Agents emerges from stealth with open standard and runtime for agents — gajesh · 2026-11-12
- From electrons to atoms: AI now drafts PCBs and CAD files for robot builds — Herodont5915 · 2026-11-12